Institutions trimmed the Magnificent Seven and bought chips. That is rotation, not an exit

Q2 2026 13F filings covering 6,371 institutional investors show 44% of them trimming Magnificent Seven holdings while 48% were net buyers of semiconductors ^[1][2][7]. The useful reading is not that professional money is backing away from AI, but that it is moving down the stack from the companies selling AI to the companies supplying it.

Start with the megacaps, because the headline number is softer than it looks. Against the 44% who trimmed, 42% increased or initiated positions ^[2][3], a net selling skew of roughly two percentage points ^[1] and about 14% of filers who did neither ^[4]. In a sample this size that is close to noise. Calling it a retreat from Microsoft, Meta and their peers ^[4] overstates a coin flip.

The chip figures are not a coin flip. 48% of filers were net buyers of semiconductors against 34.5% net sellers ^[7][8], a 13.5-point skew toward buying ^[2]. Software went the other way, with 28.2% net sellers against 26.3% net buyers ^[5][6], a 1.9-point tilt toward selling ^[3]. So the same institutions that could not decide about the platform companies were decisive about the fabs and the fabless designers. Crypto Briefing reads this as conviction in the physical infrastructure layer over the application layer ^[13], and the arithmetic supports the direction if not the drama: the chip buy skew is roughly seven times the size of the software sell skew ^[5].

Two things complicate the clean version of this story. First, the AI-themed names closest to the data center buildout drew only about 36% net buyers, including CoreWeave and Arista Networks ^[10] - twelve points below the semiconductor share ^[6]. If the thesis were purely "own the suppliers," the infrastructure names would not be lagging the chipmakers by that margin. Second, Tiger Global reduced stakes in several key tech names and also cut its Taiwan Semiconductor position ^[9], which is the opposite of the aggregate trade. Aggregates are not consensus; they are averages of disagreement.

Context matters for how much of this is thesis and how much is plumbing. Market observers have flagged July's unwinding of crowded trades as background to the positioning ^[12], and a crowded-trade unwind produces trimming that has nothing to do with a view on 2027 earnings. 13Fs are also backward-looking snapshots, so what you are reading is where the money was, not where it is.

Elsewhere in the filings, energy was the clear loser, with institutions broadly net sellers of oil holdings ^[11].

What to watch: whether the semiconductor buy skew holds in the next filing cycle or reverses as a July artifact; whether the AI infrastructure names close the twelve-point gap to chips ^[6]; and whether software's narrow selling tilt widens, given that enterprise software multiples remain elevated relative to historical norms according to the same report ^[14].